Neural Network Based Inverse Kinematics Solution Method for Origami Robots

Abstract

Inverse kinematics (IK) of origami robots is highly challenging due to their nonlinear geometry and complex folding constraints. Traditional iterative or analytical methods often suffer from high computational cost, poor convergence, and limited robustness in real-time scenarios. To address these issues, this study proposes a neural network–based IK solution framework. A dataset of 100,000 posture–joint pairs was generated through simulation, and a multilayer perceptron (MLP) was trained to approximate the nonlinear mapping from end-effector pose to joint angles. Experimental validation demonstrates that the proposed model achieves an average joint angle prediction error below 2°, representing a >40% reduction compared with conventional numerical iteration. The inference speed is approximately 20 times faster, and the convergence success rate reaches 98%, significantly surpassing baseline methods. Robustness tests under noisy inputs and boundary configurations show that prediction errors increase by less than 1°, confirming strong stability and generalization. These results indicate that the proposed neural network approach provides an efficient and reliable IK solver for origami robots, with promising applications in flexible manufacturing, space structures, and minimally invasive surgical robotics.